ICCV 2015 Papers

Although the Tree Structured Model (TSM) is proven effective for solving face detection, pose estimation and landmark localization in an unified model, its sluggish run time makes it unfavorable in practical applications, especially when dealing with cases of multiple faces. We propose the Regressive Tree Structure Model (RTSM) to improve the run-time speed and localization accuracy. The RTSM is composed of two component TSMs, the coarse TSM (c-TSM) and the refined TSM (r-TSM), and a Bilateral Support Vector Regressor (BSVR). The c-TSM is built on the low-resolution octaves of samples so that it provides coarse but fast face detection. The r-TSM is built on the mid-resolution octaves so that it can locate the landmarks on the face candidates given by the c-TSM and improve precision. The r-TSM based landmarks are used in the forward BSVR as references to locate the dense set of landmarks, which are then used in the backward BSVR to relocate the landmarks with large localization errors. The forward and backward regression goes on iteratively until convergence. The performance of the RTSM is validated on three benchmark databases, the Multi-PIE, LFPW and AFW, and compared with the latest TSM to demonstrate its efficacy.

In this work we present a method to estimate a 3D face shape from a single image. Our method is based on a cascade regression framework that directly estimates face landmarks locations in 3D. We include the knowledge that a face is a 3D object into the learning pipeline and show how this information decreases localization errors while keeping the computational time low. We predict the actual positions of the landmarks even if they are occluded due to face rotation. To support the ability of our method to reliably reconstruct 3D shapes, we introduce a simple method for head pose estimation using a single image that reaches higher accuracy than the state of the art. Comparison of 3D face landmarks localization with the available state of the art further supports the feasibility of a single-step face shape estimation. The code, trained models and our 3D annotations will be made available to the research community.

Stereo reconstruction in presence of reality faces many challenges that still need to be addressed. This paper considers reflections, which introduce incorrect matches due to the observation violating the diffuse-world assumption underlying the majority of stereo techniques. Unlike most existing work, which employ regularization or robust data terms to suppress such errors, we derive two least squares models from first principles that generalize diffuse world stereo and explicitly take reflections into account. These models are parametrized by depth, orientation and material properties, resulting in a total of up to 5 parameters per pixel that have to be estimated. Additionally large non-local interactions between viewed and reflected surface have to be taken into account. These two properties make inference of the model appear prohibitive, but we present evidence that inference is actually possible using a variant of patch match stereo.

In the past decade, Grassmann manifolds (Grassmannian) have been commonly used in mathematical formulations of many Computer Vision tasks. Averaging points on a Grassmann manifold is a very common operation in many applications including but not limited to, tracking, action recognition, video-face recognition, face recognition, etc. Computing the intrinsic/Frechet mean (FM) of a set of points on the Grassmann can be cast as finding the global optimum (if it exists) of the sum of squared geodesic distances cost function. A common approach to solve this problem involves the use of the gradient descent method. An alternative way to compute the FM is to develop a recursive/inductive definition that does not involve optimizing the aforementioned cost function. In this paper, we propose one such computationally efficient algorithm called the it Grassmann inductive Frechet mean estimator (GiFME). In developing the recursive solution to find the FM of the given set of points, GiFME exploits the fact that there is a closed form solution to find the FM of two points on the Grassmann. In the limit as the number of samples tends to infinity, we prove that GiFME converges to the FM (this is called the weak consistency result on the Grassmann manifold). Further, for the finite sample case, in the limit as the number of sample paths (trials) goes to infinity, we show that GiFME converges to the finite sample FM. Moreover, we present a bound on the geodesic distance between the estimate from GiFME and the true FM. We present several experiments on synthetic and real data sets to demonstrate the performance of GiFME in comparison to the gradient descent based (batch mode) technique. Our goal in these applications is to demonstrate the computational advantage and achieve comparable accuracy to the state-of-the-art.

We propose the Encoder-Recurrent-Decoder (ERD) model for recognition and prediction of human body pose in videos and motion capture. The ERD model is a recurrent neural network that incorporates nonlinear encoder and decoder networks before and after recurrent layers. We test instantiations of ERD architectures in the tasks of motion capture (mocap) generation, body pose labeling and body pose forecasting in videos. Our model handles mocap training data across multiple subjects and activity domains, and synthesizes novel motions while avoiding drifting for long periods of time. For human pose labeling, ERD outperforms a per frame body part detector by resolving left-right body part confusions. For video pose forecasting, ERD predicts body joint displacements across a temporal horizon of 400ms and outperforms a first order motion model based on optical flow. ERDs extend previous Long Short Term Memory (LSTM) models in the literature to jointly learn representations and their dynamics. Our experiments show such representation learning is crucial for both labeling and prediction in space-time. We find this is a distinguishing feature between the spatio-temporal visual domain in comparison to 1D text, speech or handwriting, where straightforward hard coded representations have shown excellent results when directly combined with recurrent units.

In this work we present a novel algorithm for realtime visual odometry for a monocular camera. The main idea is to develop an approach between classical feature-based visual odometry systems and modern direct dense/semi-dense methods, trying to benefit from the best attributes of both. Similar to feature-based systems, we extract information from the images, instead of working with raw image intensities as direct methods. In particular, the information extracted are the edges present in the image, while the rest of the algorithm is designed to take advantage of the structural information provided when pixels are treated as edges. Edge extraction is an efficient and higly parallelizable operation. The edge depth information extracted is dense enough to allow acceptable surface fitting, similar to modern semi-dense methods. This is a valuable attribute that feature-based odometry lacks. Experimental results show that the proposed method has similar drift than state of the art feature-based and direct methods, and is a simple algorithm that runs at realtime and can be parallelized. Finally, we have also developed an inertial aided version that successfully stabilizes an unmanned air vehicle in complex indoor environments using only a frontal camera, while running the complete solution in the embedded hardware on board the vehicle.

We present an object detector coupled with pose estimation directly in a single compact and simple model, where the detector shares extracted image features with the pose estimator. The output of the classification of each candidate window consists of both object score and likelihood map of poses. This extension introduces negligible overhead during detection so that the detector is still capable of real time operation. We evaluated the proposed approach on the problem of vehicle detection. We used existing datasets with viewpoint/pose annotation (WCVP, 3D objects, KITTI). Besides that, we collected a new traffic surveillance dataset COD20k which fills certain gaps of the existing datasets and we make it public. The experimental results show that the proposed approach is comparable with state-of-the-art approaches in terms of accuracy, but it is considerably faster - easily operating in real time (Matlab with C++ code). The source codes and the collected COD20k dataset are made public along with the paper.

In many fine-grained object recognition datasets, image orientation (left/right) might vary from sample to sample. Since handcrafted descriptors such as SIFT are not reversal invariant, the stability of image representation based on them is consequently limited. A popular solution is to augment the datasets by adding a left-right reversed copy for each original image. This strategy improves recognition accuracy to some extent, but also brings the price of almost doubled time and memory consumptions. In this paper, we present RIDE (Reversal Invariant Descriptor Enhancement) for fine-grained object recognition. RIDE is a generalized algorithm which cancels out the impact of image reversal by estimating the orientation of local descriptors, and guarantees to produce the identical representation for an image and its left-right reversed copy. Experimental results reveal the consistent accuracy gain of RIDE with various types of descriptors. We also provide insightful discussions on the working mechanism of RIDE and its generalization to other applications.

Inspired by the recent success of RGB-D cameras, we propose the enrichment of RGB data with an additional "quasi-free" modality, namely, the wireless signal (e.g., wifi or Bluetooth) emitted by individuals' cell phones, referred to as RGB-W. The received signal strength acts as a rough proxy for depth and a reliable cue on their identity. Although the measured signals are highly noisy (more than 2m average localization error), we demonstrate that the combination of visual and wireless data significantly improves the localization accuracy. We introduce a novel image-driven representation of wireless data which embeds all received signals onto a single image. We then indicate the ability of this additional data to (i) locate persons within a sparsity-driven framework and to (ii) track individuals with a new confidence measure on the data association problem. Our solution outperforms existing localization methods by a significant margin. It can be applied to the millions of currently installed RGB cameras to better analyze human behavior and offer the next generation of high-accuracy location-based services.

Hyperspectral imaging usually lack of spatial resolution due to limitations of hardware design of imaging sensors. On the contrary, latest imaging sensors capture a RGB image with resolution of multiple times larger than a hyperspectral image. In this paper, we present an algorithm to enhance and upsample the resolution of hyperspectral images. Our algorithm consists of two stages: spatial upsampling stage and spectrum substitution stage. The spatial upsampling stage is guided by a high resolution RGB image of the same scene, and the spectrum substitution stage utilizes sparse coding to locally refine the upsampled hyperspectral image through dictionary substitution. Experiments show that our algorithm is highly effective and has outperformed state-of-the-art matrix factorization based approaches.

This paper studies the problem of improving the top-1 accuracy of RGB-D object recognition. Despite of the impressive top-5 accuracies achieved by existing methods, their top-1 accuracies are not very satisfactory. The reasons are in two-fold: (1) existing similarity measures are sensitive to object pose and scale changes, as well as intra-class variations; and (2) effectively fusing RGB and depth cues is still an open problem. To address these problems, this paper first proposes a new similarity measure based on dense matching, through which objects in comparison are warped and aligned, to better tolerate variations. Towards RGB and depth fusion, we argue that a constant and golden weight doesn't exist. The two modalities have varying contributions when comparing objects from different categories. To capture such a dynamic characteristic, a group of matchers equipped with various fusion weights is constructed, to explore the responses of dense matching under different fusion configurations. All the response scores are finally merged following a learning-to-combination way, which provides quite good generalization ability in practice. The proposed approach win the best results on several public benchmarks, e.g., achieves 92.7% top-1 test accuracy on the Washington RGB-D object dataset, with a 5.1% improvement over the state-of-the-art.

Detection of elongated structures in 2D images and 3D image stacks is a critical prerequisite in many applications and Machine Learning-based approaches have recently been shown to deliver superior performance. However, these methods essentially classify individual locations and do not explicitly model the strong relationship that exists between neighboring ones. As a result, isolated erroneous responses, discontinuities, and topological errors are present in the resulting score maps. We solve this problem by projecting patches of the score map to their nearest neighbors in a set of ground truth training patches. Our algorithm induces global spatial consistency on the classifier score map and returns results that are provably geometrically consistent. We apply our algorithm to challenging datasets in four different domains and show that it compares favorably to state-of-the-art methods.

Recently, very high-dimensional feature representations, e.g., Fisher Vector, have achieved excellent performance for visual recognition and retrieval. However, these lengthy representations always cause extremely heavy computational and storage costs and even become unfeasible in some large-scale applications. A few existing techniques can transfer very high-dimensional data into binary codes, but they still require the reduced code length to be relatively long to maintain acceptable accuracies. To target a better balance between computational efficiency and accuracies, in this paper, we propose a novel embedding method called Binary Projection Bank (BPB), which can effectively reduce the very high-dimensional representations to medium-dimensional binary codes without sacrificing accuracies. Instead of using conventional single linear or bilinear projections, the proposed method learns a bank of small projections via the max-margin constraint to optimally preserve the intrinsic data similarity. We have systematically evaluated the proposed method on three datasets: Flickr 1M, ILSVR2010 and UCF101, showing competitive retrieval and recognition accuracies compared with state-of-the-art approaches, but with a significantly smaller memory footprint and lower coding complexity.

Thanks to the recent advances in computational photography and remote sensing, point clouds of buildings are becoming increasingly available, yet their processing poses various challenges. In our work, we tackle the problem of point cloud completion and editing and we approach it via inverse procedural modeling. Contrary to the previous work, our approach operates directly on the point cloud without an intermediate triangulation. Our approach consists of 1) semi-automatic segmentation of the input point cloud with segment comparison and template matching to detect repeating structures, 2) a consensus-based voting schema and a pattern extraction algorithm to discover completed terminal geometry and their patterns of usage, all encoded into a context-free grammar, and 3) an interactive editing tool where the user can create new point clouds by using procedural copy and paste operations, and smart resizing. We demonstrate our approach on editing of building models with up to 1.8M points. In our implementation, preprocessing takes up to several minutes and a single editing operation needs from one second to one minute depending on the model size and the operation type.

We consider classification problems in which the label space has structure. A common example is hierarchical label spaces, corresponding to the case where one label subsumes another (e.g., animal subsumes dog). But labels can also be mutually exclusive (e.g., dog vs cat) or unrelated (e.g., furry, carnivore). To jointly model hierarchy and exclusion relations, the notion of a HEX (hierarchy and exclusion) graph was introduced. This combined a conditional random field (CRF) with a deep neural network (DNN), resulting in state of the art results when applied to visual object classification problems where the training labels were drawn from different levels of the ImageNet hierarchy (e.g., an image might be labeled with the basic level category "dog", rather than the more specific label "husky"). In this paper, we extend the HEX model to allow for soft or probabilistic relations between labels, which is useful when there is uncertainty about the relationship between two labels (e.g., an antelope is "sort of" furry, but not to the same degree as a grizzly bear). We call our new model pHEX, for probabilistic HEX. We show that the pHEX graph can be converted to an Ising model, which allows us to use existing off-the-shelf inference methods (in contrast to the HEX method, which needed specialized inference algorithms). Experimental results show significant improvements in a number of large-scale visual object classification tasks, outperforming the previous HEX model.

Statistical shape and appearance models are often based on the accurate identification of one-to-one correspondences in a training data set. At the same time, the determination of these corresponding landmarks is the most challenging part of such methods. Hufnagel etal developed an alternative method using correspondence probabilities for a statistical shape model. We propose the use of probabilistic correspondences for statistical appearance models by incorporating appearance information into the framework. A point-based representation is employed representing the image by a set of vectors assembling position and appearances. Using probabilistic correspondences between these multi-dimensional feature vectors eliminates the need for extensive preprocessing to find corresponding landmarks and reduces the dependence of the generated model on the landmark positions. Then, a maximum a-posteriori approach is used to derive a single global optimization criterion with respect to model parameters and observation dependent parameters, that directly affects shape and appearance information of the considered structures. Model generation and fitting can be expressed by optimizing the same criterion. The developed framework describes the modeling process in a concise and flexible mathematical way and allows for additional constraints as topological regularity in the modeling process. Furthermore, it eliminates the demand for costly correspondence determination. We apply the model for segmentation and landmark identification in hand X-ray images, where segmentation information is modeled as further features in the vectorial image representation. The results demonstrate the feasibility of the model to reconstruct contours and landmarks for unseen test images. Furthermore, we apply the model for tissue classification, where a model is generated for healthy brain tissue using 2D MRI slices. Applying the model to images of stroke patients the probabilistic correspondences are used to classify between healthy and pathological structures. The results demonstrate the ability of the probabilistic model to recognize healthy and pathological tissue automatically.

We present a simple approach for producing a small number of structured visual outputs which have high recall, for a variety of tasks including monocular pose estimation and semantic scene segmentation. Current state-of-the-art approaches learn a single model and modify inference procedures to produce a small number of diverse predictions. We take the alternate route of modifying the learning procedure to directly optimize for good, high recall sequences of structured-output predictors. Our approach introduces no new parameters, naturally learns diverse predictions and is not tied to any specific structured learning or inference procedure. We leverage recent advances in the contextual submodular maximization literature to learn a sequence of predictors and empirically demonstrate the simplicity and performance of our approach on multiple challenging vision tasks including achieving state-of-the-art results on multiple predictions for monocular pose-estimation and image foreground/background segmentation.

We address the problem of recognizing a place depicted in a query image by using a large database of geo-tagged images at a city-scale. In particular, we discover features that are useful for recognizing a place in a data-driven manner, and use this knowledge to predict useful features in a query image prior to the geo-localization process. This allows us to achieve better performance while reducing the number of features. Also, for both learning to predict features and retrieving geo-tagged images from the database, we propose per-bundle vector of locally aggregated descriptors (PBVLAD), where each maximally stable region is described by a vector of locally aggregated descriptors (VLAD) on multiple scale-invariant features detected within the region. Experimental results show the proposed approach achieves a significant improvement over other baseline methods.

In this paper we address three different computer vision tasks using a single basic architecture: depth prediction, surface normal estimation, and semantic labeling. We use a multiscale convolutional network that is able to adapt easily to each task using only small modifications, regressing from the input image to the output map directly. Our method progressively refines predictions using a sequence of scales, and captures many image details without any superpixels or low-level segmentation. We achieve state-of-the-art performance on benchmarks for all three tasks.